← Previous: Preserving grilling session value · README · Next: Running parallel grilling sessions
The failure mode
- Using too dumb a model for grilling (
8:40). - Grilling demands hard judgement calls from the model: telling low-fidelity from high-fidelity questions, and figuring out which questions will prompt a stronger design (
8:40).
Two sources of model knowledge
| Source | What it is | Reliability |
|---|---|---|
| Contextual knowledge | What you passed in: files read, user prompts, tool-call results brought back from research (8:40, 9:05) | Higher |
| Parametric knowledge | What the model was trained to see and understand (9:05) | Much less reliable — but it's what grilling depends on (9:05) |
Why grilling is parametric-heavy
- Grilling leans on the model's innate understanding of systems and applications to surface ideas you hadn't considered (
9:05). - Logical consequence: anything you had considered you would already have passed in as contextual knowledge — so the value the model adds is precisely the off-the-wall suggestions and strange ideas from its parameters (
9:32). - Relying on parametric knowledge means you need a model with lots of parameters — i.e. the big frontier models, which are also top-of-the-line trained and simply more capable than smaller ones (
9:32).
The corollary most people miss
- You can use a dumber model for implementation, because by then most of the information is contextual: a detailed implementation plan plus the relevant codebase files to copy from (
9:58,10:21).
flowchart LR
G["Grilling<br/>mostly parametric"] -->|needs| F["Frontier / big model"]
I["Implementation<br/>mostly contextual"] -->|tolerates| S["Smaller, cheaper model"]